Practical details

There are several MPI-systems available. I have fetched and compiled the Open MPI-system (an older version comes with the installed RedHat-system). 

Update: May 17. There is a problem with OpenMPI and broadcasts. Use shared memory communication, so -mca btl sm,self . It works on the remote-computers and the math lab-computers. I don't know yet what happens if you try to use more than one system.

This is how you use it:

1. I have set up Open MPI in    /chalmers/sw/unsup/openmpi-1.4.1/bin   on the system (not www).  You want this path to come before the default version already installed in the system. Your LD_LIBRARY_PATH must contain /chalmers/sw/unsup/openmpi-1.4.1/lib . You find documentation on the web or under /chalmers/sw/unsup/openmpi-1.4.1/share .
2. Compile your program:
   mpicc my_prog.c  or  mpiCC my_prog.cc or  mpif90 my_prog.f90 or  mpif77 my_prog.f
3. To run on your own lab-computer (you can run several processes even if the computer only has one CPU, but you do not get true parallel computing of course, nor do you expect a speedup). In this example we run two (-n 2) processes:
   
   
   mpirun -n 2 -mca btl tcp,self ./a.out
    
   
   Using the mca-flag (MPI Component Architecture) you can control how mpirun should do certain things, e.g. what kind of communication software/hardware that should be used. -btl tcp,self (btl = byte transfer layer) informs mpirun that tcp (transmission control protocol) should be used, tcp is available on almost all systems but may not be very fast. self is nedded so that a process may communicate with itself. If your CPUs can communicate through shared memory, you can try -mca btl sm,self instead. On a more expensive system a high performance network, like Infiniband may be available.
4. To run on several computers, create a hostfile (named hostfile in the example) containing a list of names of computers, one name per row. Suppose you want to run on one of our parallel machines, having four cpus. Login on such a machine and give the command hostname, this gives you the full Internet domain name of the parallel computer.
   
   
   mpirun -n 2 -hostfile hostfile -mca btl tcp,self ./a.out